Integrating Geophysics and Machine Learning for Soil Compaction Estimation in Agriculture

Soil compaction is a significant factor affecting crop yield; however, traditional assessment methods are labor-intensive and costly. To address this issue, rapid and cost-effective estimation techniques are needed. Geophysical methods are commonly used to predict various soil parameters, yet assessing parameters related to soil compaction remains challenging. Thus, integrating geophysical methods with machine learning presents a promising alternative to pedophysical models. This study examined how machine learning techniques — such as neural networks, support vector machines, and extreme gradient boosting — can predict penetration resistance, a key indicator of soil compaction. The models were developed using geophysical data from Geonics EM38 and Veris 3100 scanners, along with soil texture-related coefficients. Penetration resistance was measured at various soil depths ranging from the surface to 0.5 m. The most accurate models were achieved with the support vector machine algorithm for penetration resistance analyzed in soil layers from 0 to 0.4 m, utilizing soil apparent electrical conductivity and magnetic susceptibility measured with the EM38 scanner at a depth of 0.5 m, electrical conductivity from the Veris 3100 scanner, and the soil texture coefficient as independent variables. For these models, the mean percentage error (MAPE) for the test data fell between 12% and 23%, while the correlation coefficient R ranged from 0.66 to 0.79. The resulting models can be practically applied to plan tillage at various depths, optimizing fuel consumption without compromising soil conditions necessary for plant root system development.